Facet-dependent oxygen evolution on IrO2 from machine-learned potential-driven atomistic models

Abstract

Hydrogen production by acidic water electrolysis is limited by the anodic oxygen evolution reaction (OER), for which IrO₂ remains the state-of-the-art catalyst as it combines low overpotential with comparatively high activity. However, the performance of this catalyst is strongly facet dependent, and the atomistic origin of the activity gap across different surface orientations remains incompletely understood due to the complex, discrete interplay of interfacial solvation, redox chemistry, and adsorbate energetics coupled under electrochemical conditions. Here, we present a unified thermodynamics and electronic structure analysis of the OER on IrO₂(110) and IrO₂(101) surfaces that explicitly accounts for the facet-dependent OER activity. We first evaluated the OER free energy profiles using the CHE model with implicit solvation; however, the facet-dependent OER trend was not clearly captured. Subsequently, we incorporate explicit water solvation using machine-learned molecular dynamics simulations (MLMD) to obtain equilibrated IrO₂/H₂O structures. Using water-ligated Ir sites as representative catalytic sites, we compute adsorption free energies for the OER intermediates and construct free energy diagrams. The reaction pathway determined by the explicit solvation model identifies *O to *OOH transition as the potential-determining step on both facets, with IrO₂(110) requiring a substantially lower overpotential than IrO₂(101). Electronic structure analyses, including work function, spin-resolved projected density of states, and orbital-resolved COHP, reveal that the (101) facet stabilizes the Ir–O state more strongly through the facet-dependent redistribution of Ir (e_g/t_2g) hybridization with O(2p) states, consistent with larger charge transfer and stronger Ir–O bonding. This stronger oxo stabilization increases the corresponding energy requirement for O–O bond formation, rationalizing the higher overpotential of IrO₂(101). Collectively, these results show that the explicit solvation model driven by machine learning potential provide a better mechanistic analysis of the OER activity on the IrO₂ facets.